🩸 @red_lili — Graphic Madness & High-Contrast Anime Style for Anima (DiT)

A surgical, hyper-compact style LoRA capturing an intense, unhinged, high-contrast aesthetic with expressive ringed eyes, bold lineart, and dramatic cel-shading.

---

🔬 Why this LoRA is built different (3 MB & Zero-Leakage)

Unlike standard brute-force LoRAs that weigh 200MB+ and ruin base anatomy, this model was trained using [Anima-training-framework] — a bespoke architecture-aware trainer leveraging mechanistic interpretability and Function-Space Prior regularization.

- ⚡ Ultra-Compact (~3 MB): Only 1.4M parameters across 5 core DiT blocks (Rank 6 MLP, Rank 2 Modulation, Cross-Attention). Zero bloat.

- 🔒 True Zero-Leakage: Without the trigger word, the base model remains 100% untampered preservation_drift < 0.006).

- 🎨 Orthogonal Style / Prompt Adherence: The LoRA learns how to render, not what to render. It naturally respects medium modifiers like sketch, monochrome, crosshatching, or full color without fighting your prompt.

- 📐 Trained on 1280x1280 native resolution.

---

⚙️ Recommended Settings

- Base Model: Anima anima_baseV10)

- Trigger Word: @red_lili

- LoRA Weight: 0.8 – 1.0

- CFG Scale: 4.0 – 6.0

- Sampler: dpmpp_2m_sde_gpu or euler (20–30 steps)

---

💡 Prompting Tips

- Core Style: @red_lili, crazy smile, ringed eyes, heavy blush, blood splatter, red background, high contrast, dynamic angle

- Sketch / Manga Mode: @red_lili, sketch, rough sketch, monochrome, crosshatching, spot color

- Color Palettes: Works great with limited palettes or bold vibrant backgrounds.

---

🛠️ Trainer Source Code & Research:

Trained with [Anima Training Framework on GitHub] — an architecture-aware DiT trainer with selective layer-targeting and L2 function-space anchoring.